Objective: To quantify the short-duration levodopa response using high resolution data derived from a wrist sensor and processed with advanced deep learning methods

Background: The fluctuating and highly heterogeneous nature of Parkinson’s disease necessitates treatment adjustments over time; however, the lack of objective, longitudinal data in Parkinson’s monitoring is a major limitation to the timely personalization of care and patient outcomes including quality of life. Currently care teams rely on brief, clinical appointments and subjective, labor-intensive tools such as rating scores or patient motor diaries to gauge treatment effects.

Method: In this investigation we deployed a novel combination of continuous monitoring and machine learning techniques to passively capture Parkinson’s disease motor fluctuations in 91 patients living with Parkinson’s in Germany. Motion data collected from a single wrist sensor provided input for a fully convolutional neural network (FCN) trained to continuously detect PD motor states and their severity. Various wrist sensor models that collect accelerometer and gyroscope data were used in the device-agnostic study design. The timing of medication intake was recorded, and longitudinal visualization and hierarchical clustering analyses applied to model output.

Results: In total 285 levodopa cycles were captured among the cohort of 91 Parkinson’s patients (59% male, mean age 68 +/- 10 years (+/- SD), disease duration 9 +/- 6 years). Hierarchical clustering classified single levodopa intakes into one of three levodopa response clusters based primarily on minimum and maximum severity of dyskinesia and bradykinesia on the 9 item scale used to label training data for the deep learning algorithm.

Conclusion: This technology-enabled approach yielded multidimensional insights into the short-duration levodopa response. Continuous data and its processing with AI methods were able to capture longitudinal motor state and treatment response information on the time frame of a single levodopa cycle at the level of an individual patient. Our results suggest that unintrusive sensors and machine learning methods in Parkinson’s disease care can provide continuous, objective, and individual-level data for monitoring treatment response and be used to infer information about treatment effect size and timing, and need for further optimization of Parkinson’s care.

« Back to 2023 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/continuous-algorithmic-insights-into-pd-motor-symptoms-and-individual-treatment-response/